Orthogonal Transformations

Math 217 Professor Karen Smith (c)2015 UM Math Dept

licensed under a Creative Commons
By-NC-SA 4.0 International License.

T
Definition: A linear transformation Rn −→ Rn is orthogonal if |T (~x)| = |~x| for all ~x ∈ Rn .
T
Theorem: If Rn −→ Rn is orthogonal, then ~x · ~y = T ~x · T ~y for all vectors ~x and ~y in Rn .

A. Last time you proved:

1. An orthogonal transformation is an isomorphism.

2. The inverse of an orthogonal transformation is also orthogonal.

3. The composition of orthogonal transformations is orthogonal.

Discuss with your table the geometric intuition of each of these statements. Why do they make
sense, for example, in R3 .
 
2 2 a c
B. Suppose T : R → R is given by left multiplication by A = . Assuming T is orthogonal,
b d
illustrate an example of such a T by showing what it does to a unit square. Label your picture
with a, b, c, d. What do you think det A can be?
Solution note: Your picture should show a square of sides 1 with vertices (0, 0), (a, b), (c, d)
and (a + c, b + d). The determinant of A is either 1 or -1, since | det A| is the area of
this square.

C. Definition: An n × n matrix is orthogonal if its columns are orthonormal.

 
cosθ −sinθ
1. Is A = orthogonal? Is the map “multiplication by A” orthogonal? Why?
sinθ cosθ

Solution note: Yes, the columns are orthonormal since the (cosθ)2 + (−sinθ)2 = 1,
and (cosθ)(−sinθ) + (sinθ)(cosθ) = 0.. Multiplication by A is rotation through an
angle of θ (counterclockwise), so it preserves lengths and hence is orthogonal.

 
3/5 4/5 0
2. Is B = −4/5 3/5 0 orthogonal? Is B T orthogonal? Is the map “multiplication by B”
0 0 1
orthogonal? Why? Describe it geometrically.
Solution note: Yes, again, just check that the dot product of the columns is either 1
or 0, depending on whether we dot a column with itself or a different column. Same
is true for B T . Multiplication by A is orthogonal...actually it is a rotation around
the e3 -axis.
 T
D. Prove the following Theorem: Let Rn −→ Rn be the linear transformation T (~x) = A~x, where
A is an n × n matrix. Then T is orthogonal if and only if the matrix A has orthonormal columns.
[Hint: Scaffold first. What are the two things you need to show?]
Solution note: We need to show two things:
1). If T is orthogonal, then A has orthonormal columns. Assume T is orthogonal.
The columns of A are [T (~e1 ) T (~e2 ) . . . T (~en )]. To check they are orthonormal, we
need two things: T (~ei ) · T (~ei ) = 1 for all i, and T (~ei ) · T (~ej ) = 0 if i 6= j. For the
first, note
T (~ei ) · T (~ei ) = ||T (~ei )||2 = ||~ei ||2 = 1,
with the first equality coming from the definition of the length of the vector and the
second coming from the definition of T being orthogonal. For the second, T (~ei ) ·
T (~ej ) = ~ei · ~ej by the Theorem above. This is zero since the standard basis is
orthonormal.
2). Assume A has orthonormal columns. We need to show that for any ~x ∈ Rn ,
||T (~x)|| = ||~x||. Since the length is always a non-negative number, it suffices to show
||T (~x)||2 = ||~x||2 . That is, it suffices to show T (~x) · T (~x) = ~x · ~x. For, this we take an
arbitrary ~x and write it in the basis {~e1 , . . . , ~en }. Note that
X
~x · ~x = (x1~e1 + · · · + xn~en ) · (x1~e1 + · · · + xn~en ) = (xi~ei ) · (xj ~ej ) = x21 + · · · + x2n .
ij

here the third equality is using some basic properties of dot product (like ”foil”), and
the third equality is using the fact that the ~ei are orthonormal so that (xi~ei )·(xj ~ej ) = 0
if i 6= j. On the other hand, we also have

T (~x) · T (~x) =T (x1~e1 + · · · + xn~en ) · T (x1~e1 + · · · + xn~en )

X
= (xi T (~ei )) · (xj T (~ej ))
ij
X
= xi xj T (~ei ) · T (~ej ) = x21 + · · · + x2n
ij

with the last equality coming from the fact that the T (~ei )’s are the columns of A and
hence orthonormal. QED.

E. With A and B as in Problem C, compute AT A and B T B. Is AT orthogonal? Compute AT A.

What do you notice?
Solution note: The transposes of the orthogonal matrices A and B are orthogonal.
Also AT A = I2 and B T B = I3 . In general, it is true that the transpose of an
othogonal matrix is orthogonal AND that the inverse of an orthogonal matrix is its
transpose.

F. Prove that if M is an orthogonal matrix, then M −1 = M T . [Hint: write M as a row of columns

and M T as a column of rows. Compute M T · M . ]
  T
~v1
~v T 
 2
Solution note: Let M = [~v1 ~v2 ... ~vn ] where each ~vi ∈ Rn . Then M T =  . .
 .. 
~vnT
So using block multiplication, M T M is the matrix whose ij=th entry is the product
~viT ~vj , which is ~vi · ~vj . Since the ~vi ’s are orthonormal, this is the identity matrix.

G. Prove that the rows of an orthogonal matrix are also orthonormal. [Hint: don’t forget that
the rows of A are the columns of AT . Remember that the inverse of an orthogonal map is also
orthogonal.]
Solution note: Say A is orthogonal. Then the map TA is orthogonal. Hence its inverse
is orthogonal, and so the matrix of the inverse, which is A−1 is orthogonal. By the
previous problem, we know also that A−1 = AT is orthogonal. So since the columns
of AT are orthonormal, which means the rows of A are orthonormal.
H. Prove the very important Theorem on the first page saying that orthogonal transformations
preserve dot products. Why does this tell us that orthogonal transformations preserve angles?
[Hint: consider x + y.]
Solution note: Assume T is orthogonal. So ||T (x + y)|| = ||(x + y)|| for all x, y ∈ Rn ,
by definition of orthogonal. Hence also

||T (x + y)||2 = ||(x + y)||2

and so T (x + y) · T (x + y) = (x + y) · (x + y). Using linearity of T , we have

(T (x) + T (y)) · (T (x) + T (y)) = (x + y) · (x + y).

Now we expand both sides using basic properties of dot product:

(T (x) + T (y)) · (T (x) + T (y)) = (T (x) · (T (x)) + 2T (x) · T (y) + (T (y) · T (y))
= ||T (x)||2 + 2T (x) · T (y) + ||T (y)||2 .

And this is equal to

(x + y) · (x + y) = x · x + 2x · y + y · y = ||x||2 + 2x · y + ||y||2 .

That is,
||T (x)||2 + 2 T (x) · T (y) + ||T (y)||2 = ||x||2 + 2 x · y + ||y||2 .
Because T is orthogonal, ||T (x)|| = ||x|| and ||T (y)|| = ||y||, so we can cancel these
from both sides to get 2T (x)·T (y) = 2x·y. Dividing by 2, we see that T (x)·T (y) = x·y.
The proof is complete.

I. An Application of QR factorization. Suppose that we are given a system of n linear equations

in n unknowns: A~x = ~b.
1. Assuming A is invertible, express the solutions to this system in terms of A−1 and ~b.
2. Assume that A = QR is the QR-factorization of A. What does this mean? What are the
sizes of the matrices in this case?
 3. What happens if we multiply both sides of the equation A~x = ~b by QT ? How might this
simplify the problem of finding the solution to this system?
 
4 4/5 −11/5
4. Suppose that A = 0 −10 1 . Applying the Gram-Schmidt process to the columns
3 −3/5 23/5
 
4/5 0 −3/5
of this matrix we get  0 −1 0 . Find the QR factorization of A.
3/5 0 4/5

Solution
 note:
 We have A = QR where Q is the orthogonal matrix above and R is
5 1 1
0 10 −1.
0 0 5

5. Use your QR factorization to quickly solve the system A~x = [0 0 25]T without row
reducing!.

Solution note: To solve A~x = QR~x = ~b, multiply both sides by Q−1 , which
 is Q
T

15
since Q is orthogonal. We have the equivalent system R~x = QT ~b =  0 . This is
20
easy to solve because R is upper triangular. We get z = 4 from the bottom row. Then
the second row gives 10y − 4 = 0, so y = 2/5. The top row gives 5x + 2/5 + 4 = 15,
so x = 53/5.

G. TRUE OR FALSE. Justify. In all problems, T denotes a linear transformation from Rn to itself,
and A is its matrix in the standard basis.

1. If T is orthogonal, then x · y = T x · T y for all vectors x and y in Rn .

2. If T sends every pair of orthogonal vectors to another pair of orthogonal vectors, then T is
orthogonal.

3. If T is orthogonal, then T is invertible.

4. An orthogonal projection is orthogonal.

5. If A is the matrix of an orthogonal transformation T , then the columns of A are orthonormal.

6. The transpose of an orthogonal matrix is orthogonal.

7. The product of two orthogonal matrices (of the same size) is orthogonal.

8. If A is the matrix of an orthogonal transformation T , then AAT is the identity matrix.

9. If A−1 = AT , then A is the matrix of an orthogonal transformation of Rn .